MLperf power numners

srikgurrapu · July 15, 2021, 1:47am

Hello,

I’ve a question about the power measurements at: v1.0 Results | MLCommons

Test ID: 1.0-633 NVIDIA
NVIDIA Jetson AGX Xavier 32GB (MaxQ, TensorRT)

What is this MAXQ? I assume it’s some power saving mode but I couldn’t find documentation about this.

In this test case, inference times and power consumption numbers are given but there are no corresponding TOPS. What TOPS are used in this calculation?

Thanks for any advice?

Best regards,
Srik Gurrapu

AastaLLL · July 15, 2021, 3:29am

Hi,

Thanks for your question.

We are double-checking this internally.
Will get back to you soon.

AastaLLL · July 27, 2021, 6:33am

Hi,

Sorry for the late.

MAXQ is a power mode defined in the MLPerf sample.
You can find some information below:

github.com

mlcommons/inference_results_v1.0/blob/master/closed/NVIDIA/configs/3d-unet/Offline/config.json#L293


      
                  }
              },
              "gpu_offline_expected_qps": 230,
              "numa_config": "3:0-15&2:16-31&1:32-47&0:48-63&7:64-79&6:80-95&5:96-111&4:112-127",
              "scales": {
                  "A100-PCIex8": {
                      "gpu_offline_expected_qps": 0.58
                  }
              }
          },
          "AGX_Xavier": {
              "config_ver": {
                  "high_accuracy": {},
                  "maxq": {
                      "gpu_offline_expected_qps": 2.1
                  },
                  "maxq_high_accuracy": {
                      "gpu_offline_expected_qps": 2.1
                  }
              },
              "extends": [

The performance is measured in time.
You can find the detailed score for Xavier of each model below:

Thanks.

srikgurrapu · September 20, 2021, 10:26pm

Thank you for the reply. The follow-up question i have is this.

The power number for

Test ID: 1.0-633 NVIDIA
NVIDIA Jetson AGX Xavier 32GB (MaxQ, TensorRT)
NVIDIA Carmel (ARMv8.2) 1 NVIDIA AGX Xavier

Resnet power: 25.20W
SSD-small power: 23.79W
SSD-large: 23.20 W

Could you elaborate on what does this power include? I assume it is SoC plus DDR memories needed to run inference but wanted to confirm,

Also, when I look at the log files below, the power is 17W - which is different from the above link.

github.com

mlcommons/inference_results_v1.0/blob/master/closed/NVIDIA/results/AGX_Xavier_TRT_MaxQ/ssd-resnet34/SingleStream/performance/power/ptd_logs.txt

Yokogawa ethernet/USB interface specified.
Selected power meter 'Yokogawa WT330:1-Channel' from wt310.cpp
  ****************************************************************************
                      ***********************************                     
                               SPEC PTDaemon Tool                             
                        Version 1.9.2-3976349f-20201208                     
                      ***********************************                     
                     Licensed Materials - Property of SPEC
     Copyright 2006-2020 Standard Performance Evaluation Corporation (SPEC)
                              All Rights Reserved.
  For use with benchmark products from SPEC and authorized organizations only.
  ****************************************************************************


Redirecting data output to file C:\logs_ptdeamon.txt
Calculated PTD CRC: 0x3976349f, 1296896
03-18-2021 00:44:46.079: Attempting to connect to measurement device type 52...
03-18-2021 00:44:46.082: Yokogawa TMCTL library loaded.
03-18-2021 00:44:46.082: Yokogawa ethernet mode selected
03-18-2021 00:44:46.708: Analyzer identity response of 33 bytes: YOKOGAWA,WT333E,C2VL22012V,F1.03

This file has been truncated. show original

Thank you for any clarification.

kayccc · September 29, 2021, 2:29am

Hi srikgurrapu,

Please help to open a new topic. Thanks

Topic		Replies	Views
How can I test the NVIDIA website to give MLperf score using the development board in my hand? Jetson Xavier NX ml	8	763	April 17, 2024
performance of AGX Jetson AGX Xavier	7	1074	December 26, 2018
MAXn power mode in production? Jetson AGX Xavier power , performance	20	3316	April 6, 2023
How to get reliable GPU clock rate on Jetson AGX Xavier? Jetson AGX Xavier tensorrt	2	671	October 18, 2021
Orin mlperf result Jetson AGX Orin performance	5	1618	May 25, 2022
Maximum power consumption bounds for Orin AGX Jetson AGX Orin power	28	191	March 15, 2025
Power consumption difference between AGX and NX Jetson AGX Xavier power	17	2988	October 18, 2021
Inference Speed Jetson Xavier NX pytorch	6	868	April 12, 2023
Xavier 10G pcie switch bandwidth is low，[ksoftirQD /0] load Is too high Jetson AGX Xavier pcie , ethernet	15	1029	July 28, 2022
Xavier Power consumption information (Solved) Jetson AGX Xavier	17	15184	October 18, 2021

MLperf power numners

Related topics